main report - Did SARS-CoV-2 first infect humans via zoonotic spillover or a lab leak

Did SARS-CoV-2 first infect humans via zoonotic spillover or a lab leak?

1. The answer

Most likely natural zoonotic spillover. My bottom-line estimate is ~83% zoonosis vs ~17% research-related (lab leak), with a defensible range of ~80–88% zoonosis — this figure is the report’s own judgment, not a model output. Of the ~17% research-related, almost all is the unmodified-collection scenario (H-42 - Research-related leak of an unmodified naturally-evolved virus (zoonotic collection)) — a wild virus that reached a human through collecting or handling it. Deliberate engineering (H-43 - Research-related origin of a laboratory-manipulated virus (engineered lab leak)) is close to ruled out.

Three things carry this. First, the origin cluster HC-1 - Origin of SARS-CoV-2 — natural zoonosis vs research-related incident prices the question directly: model posterior natural zoonosis (H-41 - Natural zoonotic spillover with no research involvement) 0.776, engineered leak (H-43 …) 0.043, residual 0.019 (run runner/run.py). Second, and largely independently, the market cluster HC-3 - Causal role of the Huanan Seafood Market in the emergence finds the Huanan market was the true animal→human spillover site (H-14 - The market environmental pattern reflects wildlife-associated emergence at the Huanan market) 0.740, not a mere downstream amplifier (H-15 - The market may have been an amplification venue seeded by infected humans, not necessarily the origin) 0.184 — and a spillover at the market is zoonosis by definition, whereas a lab leak predicts amplifier. HC-1’s genome/base-rate evidence and HC-3’s geospatial/metagenomic evidence point the same way on different data, so the combined confidence sits above HC-1’s 0.776 alone — hence my ~83%. Third, the genome evidence (the closest known natural relative, BANAL-52, CG-11 - HC-1 joint over O-68+O-69+O-70+O-71+O-72+O-73) actively suppresses the engineered branch.

What sits outside the number. The lab-leak probability rests almost entirely on one prior factor: f_wuhan, the “why did it start in Wuhan, home of the world’s leading bat-coronavirus lab?” coincidence. The model set it to a moderate 5 (defensible range 2–80). Push it to the top of that range and the answer flips: --set HC-1:f_wuhan=80 gives posterior natural H-41 … 0.199, research bloc ~0.80. So my 80–88% is conditional on f_wuhan being low-to-moderate; the full uncertainty, including that unresolved dispute, is much wider than 80–88%. Also outside the number: a hybrid origin (a lab-collected or lab-modified natural virus that then spilled over naturally) that every genome and market signal would miss, hiding in the residual H-44 - The origin is something not listed here; and the fact that both object-level zoonosis signals rest on two contested China-collected datasets re-analysed to opposite conclusions.

I would bet ~4:1 on zoonosis given the model’s moderate f_wuhan, and take no odds on f_wuhan itself — the real crux.

Entry points (start here): Analysis of HC-1 - Origin of SARS-CoV-2 — natural zoonosis vs research-related incident · Analysis of HC-3 - Causal role of the Huanan Seafood Market in the emergence · Analysis of HC-2 - Number of independent introductions of SARS-CoV-2 into humans · H-41 - Natural zoonotic spillover with no research involvement · H-14 - The market environmental pattern reflects wildlife-associated emergence at the Huanan market · CG-7 - HC-3 joint over O-1+O-2+O-7+O-30+O-31+O-32 · CG-11 - HC-1 joint over O-68+O-69+O-70+O-71+O-72+O-73 · CG-1 - HC-1 joint over O-35+O-36+O-37+O-38 · D-1 - Huanan Market environmental sample set (China CDC, Jan-Mar 2020) · D-2 - Early Wuhan case line-list and residential geolocations

2. What the analysis found

HC-1 (origin) — the direct answer. This cluster (review: Analysis of HC-1 - Origin of SARS-CoV-2 — natural zoonosis vs research-related incident) splits the origin into four exclusive histories and moved prior [0.809, 0.120, 0.047, 0.024] to posterior [0.776, 0.162, 0.043, 0.019] (natural / unmodified-leak / engineered / residual). The prior is a location-free base rate — natural zoonosis dominates because SARS-1, MERS and essentially every recent novel-pathogen pandemic was natural — times circumstantial factors, chiefly f_wuhan. The evidence raised the research bloc slightly (17%→~20%), but note which member: the DEFUSE proposal (CG-1 - HC-1 joint over O-35+O-36+O-37+O-38, a 2018 Wuhan proposal to insert exactly the furin-cleavage site later seen) and the FBI/DOE lab-lean intelligence lifted the bloc, while genome evidence pushed engineering down. Because natural spillover H-41 … and an unmodified collection leak H-42 … share an identical natural genome, all that genome evidence leaves H-42 untouched — so the research probability piled into the unmodified-leak member, the one the genome cannot rebut.

HC-3 (market) — the independent corroboration. This cluster (review: Analysis of HC-3 - Causal role of the Huanan Seafood Market in the emergence) started near-balanced (share_origin = 0.52, a deliberate coin-flip) and updated hard to posterior [0.740, 0.184, 0.048, 0.028]. Two items did it. The geospatial pattern (CG-7 - HC-3 joint over O-1+O-2+O-7+O-30+O-31+O-32, on D-2 - Early Wuhan case line-list and residential geolocations): early cases cluster on the market against a population-density null, both viral lineages centre there, and market-unlinked cases live even closer to it than linked ones — ~2.5× better explained by a causally-central market than by a detection artifact, collapsing the artifact reading H-19 - The Huanan-market early-case clustering is largely an ascertainment-bias artifact, not evidence of a market spillover to ~5%. The metagenomic pattern (CG-8 - HC-3 joint over O-17+O-18+O-19+O-20+O-28+O-29+O-33+O-40+O-41+O-42+O-50+O-51, on D-1 - Huanan Market environmental sample set (China CDC, Jan-Mar 2020)): viral positives concentrate ~87.5% in the wildlife wing, co-located with susceptible-mammal DNA — this gives origin H-14 … its edge over amplifier H-15 …, but only a ~1.4× one.

The weighing (my judgment). HC-1 alone gives ~78% zoonosis but cannot separate natural spillover from an unmodified leak — both imply a natural genome, so the genome evidence is silent between them. HC-3 breaks exactly that tie from independent data: it says the animal→human jump physically happened at the market (H-14 … 0.740), which is zoonosis and which a lab leak does not predict. Because HC-3’s evidence base (case geography, environmental swabs) is separate from HC-1’s (genome, base rates), the two agreeing lifts the combined estimate above 0.776 — my ~83%. Caveat, priced not multiplied: HC-1 and HC-3 share a depends_on — f_wuhan (“why Wuhan/the market?”) is partly the same question HC-3 resolves — so I do not multiply 0.776 × 0.740; I treat HC-3 as retiring the benign-vs-lab reading of the Wuhan coincidence, worth a few points, not an independent factor.

HC-2 (number of introductions) — bears little. This cluster (review: Analysis of HC-2 - Number of independent introductions of SARS-CoV-2 into humans) asks whether early lineages A/B came from one spillover or ≥2. The “≥2 introductions” claim is often cited as strong pro-zoonosis (two natural spillovers are near-impossible under a one-off leak). But the analysis found it weak: posterior single introduction H-16 - SARS-CoV-2 early lineages A and B are better explained by a single introduction than by two 0.736 vs ≥2 H-5 - Early A-B diversity reflects two or more separate zoonotic introductions, not one 0.264, because Pekar’s two-introduction signal (A-3 - Single introduction almost never reproduces the two-lineage two-mutation topology, favoring two or more introductions) and Weissman’s reversal of it (A-35 - Correcting Pekar 2022 asymmetric conditionalization reverses its Bayes factor to at least 4.4 favouring a single introduction) nearly cancel on the same dataset. Crucially, “single introduction” is the zoonotic null — it does not favour a lab leak. So HC-2 withholds a pro-zoonosis signal rather than supplying a pro-lab one — roughly neutral here.

3. What the answer hangs on

1. f_wuhan — the master crux. Nearly the entire circumstantial lab-tilt rides on this one factor, which has no evidence edge. Priced: at the model’s 5, research bloc ~20%; --set HC-1:f_wuhan=1 (zero the coincidence) → natural H-41 … 0.928, research bloc ~5%; --set HC-1:f_wuhan=80 → natural 0.199, research bloc ~0.80 — a full flip to lab-majority. It is the unresolved Worobey/Pekar-vs-Weissman question of whether spillover probability tracks raw wildlife-trade volume (tiny in Wuhan) or human population/urbanisation/detectability (huge in an 11M city with a live-mammal market) — this one number, not any datum, governs the top line.
2. The raccoon-dog co-location dispute — HC-3’s H-14-vs-H-15 lever, “the single most contestable number in the analysis” (CG-8’s ~1.4×). Crits-Christoph (A-48 - Concentration of viral positivity at the exact stall selling susceptible mammals raises the probability of an infected-animal source but does not prove infection, A-49 - The reported negative correlation between wildlife mtDNA and viral RNA is largely a sampling-design artifact, not evidence against a wildlife source) reads the stall-level co-location of virus with susceptible-mammal DNA as a wildlife-source signal; Bloom (A-54 - SARS-CoV-2 tracking fish and livestock rather than raccoon dogs is the pattern expected from human deposition, undercutting the specific raccoon-dog-shedding claim) reads the same bytes as human deposition; and because samples were taken ~1 month into widespread human transmission (A-55 - Because the samples were collected a month into widespread human transmission, their viral-animal co-location is uninformative about the outbreak source in either direction), co-location is a weak instrument either way. It moves origin-vs-amplifier only within zoonosis, so barely touches the top-line — but it is where HC-3’s lift is softest.
3. The origin-vs-amplifier prior share_origin. Priced: --set HC-3:share_origin=0.2 (amplifier-favouring prior) still recovers origin H-14 … 0.447 ≈ amplifier 0.445, crushing the artifact reading to ~7%; =0.8 gives 0.875. The Wuhan evidence robustly rejects “the market was not causally special,” but origin-vs-amplifier stays prior-sensitive.

Highest-value missing information (from the reviews, re-ordered by effect on the main answer): (1) a resolved value / narrow band for f_wuhan — does not exist (the field’s open crux); the largest single mover. (2) Independent sampling of the market’s live-wildlife stalls before the animals were cleared, 1 Jan 2020 — does not exist; the datum the H-14/H-15 crux turns on. (3) Any genuine discriminator between natural spillover and an unmodified leak — does not exist. (4) The un-released China-CDC D-1 metadata and a bias-free case line-list — exists, inaccessible / baked-in.

4. What this does not cover

Scope. The question was taken as: how did the virus first reach humans — natural spillover (market as leading venue) vs a research-related incident. All three in-model histories assume a clean origin; a fully hybrid one is not well-captured. The most consequential gap is the residual H-44 … (~1.9%): a lab-collected or lab-modified natural virus that then spilled over naturally would show a natural genome and a genuine market spillover — every signal here would miss it. Its low number is a guess, and the risk is out of proportion to it.

Reliance on two contested Chinese-collected datasets. The whole object-level zoonosis case rests on D-2 … (case geolocations) and D-1 … (market swabs, collected after the animals were removed), each a single motivated witness re-analysed to opposite conclusions; if either is compromised, all its observations move together.

Ingestion exclusions. 91 sources were cut to 23 at a deliberately low trust baseline (0.50), so the contested-but-central middle (Pekar, Worobey, Crits-Christoph/Bloom, Andersen, Bruttel, the Bayesian analyses) entered as discounted both-sides includes rather than being screened out (see agent-notes/curation.md). One flagged real gap: no primary source for the base rate “a novel pathogen’s first-detected cluster centres on a market regardless of true origin.”

What the debate performed rather than settled. HC-2’s near-cancellation is two competent re-analyses of one Pekar dataset pointing opposite ways — a standoff, not a wash of evidence; HC-3’s H-14/H-15 edge is likewise contested, not resolved.

External consensus (comparison, not input — the report’s own reading). The debate’s two expert judges both ruled for zoonosis, and the scientific mainstream (Worobey, Pekar, WHO SAGO) leans zoonosis. Yet the six Bayesian analyses of this same evidence spanned ~23 orders of magnitude, from Rootclaim (~89% lab) and Levin (~14,900:1 lab) at one pole to strong-zoonosis at the other. This analysis lands inside the zoonosis-leaning majority (~83%), but its contribution is the decomposition: it shows the lab case rests almost entirely on the single f_wuhan location factor (zero it and the research bloc collapses to ~5%), and rules deliberate engineering out via the BANAL/genome evidence — pinning down the one number a reader must overturn.